Skip to content

Commit

Permalink
[VP9] [Encode] Fill padding with gmm CpuBlt on MTL
Browse files Browse the repository at this point in the history 
Fill the padding for unaligned height input with gmmlib CpuBlt()
for VP9 encoding on MTL.
  • Loading branch information
@zlma7001
zlma7001 committed Sep 11, 2023
1 parent d8e2f9d commit 1c75882
Show file tree
Hide file tree
Showing 2 changed files with 108 additions and 20 deletions.
126 changes: 107 additions & 19 deletions media_softlet/agnostic/common/codec/hal/enc/vp9/packet/encode_vp9_vdenc_packet.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -791,40 +791,128 @@ void Vp9VdencPkt::fill_pad_with_value(PMOS_SURFACE psSurface, uint32_t real_heig
return;
}

if (psSurface->OsResource.TileType == MOS_TILE_INVALID)
{
return;
}

if (psSurface->Format == Format_NV12 || psSurface->Format == Format_P010)
{
uint32_t pitch = psSurface->dwPitch;
uint32_t UVPlaneOffset = psSurface->UPlaneOffset.iSurfaceOffset;
uint32_t YPlaneOffset = psSurface->dwOffset;
uint32_t pad_rows = aligned_height - real_height;
uint32_t y_plane_size = pitch * real_height;
uint32_t uv_plane_size = pitch * real_height / 2;

MOS_LOCK_PARAMS lockFlags;
MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS));
lockFlags.WriteOnly = 1;

uint8_t *src_data = (uint8_t *)m_osInterface->pfnLockResource(m_osInterface, &(psSurface->OsResource), &lockFlags);
ENCODE_CHK_NULL_NO_STATUS_RETURN(src_data);
// padding for the linear format buffer.
if (psSurface->OsResource.TileType == MOS_TILE_LINEAR)
{
uint8_t *src_data = (uint8_t *)m_osInterface->pfnLockResource(m_osInterface, &(psSurface->OsResource), &lockFlags);

uint32_t pad_rows = aligned_height - real_height;
if (!src_data)
return;

uint8_t *src_data_y = src_data + YPlaneOffset;
uint32_t y_plane_size = pitch * real_height;
uint8_t *src_data_y_end = src_data_y + y_plane_size;
uint32_t y_pitch = pitch;
for (uint32_t i = 0; i < pad_rows; i++)
{
MOS_SecureMemcpy(src_data_y_end + i * y_pitch, y_pitch, src_data_y_end - y_pitch, y_pitch);
}
uint8_t *src_data_y = src_data + YPlaneOffset;
uint8_t *src_data_y_end = src_data_y + y_plane_size;
for (uint32_t i = 0; i < pad_rows; i++)
{
MOS_SecureMemcpy(src_data_y_end + i * pitch, pitch, src_data_y_end - pitch, pitch);
}

uint8_t *src_data_uv = src_data + UVPlaneOffset;
uint32_t uv_plane_size = pitch * real_height / 2;
uint8_t *src_data_uv_end = src_data_uv + uv_plane_size;
uint32_t uv_pitch = pitch / 2;
for (uint32_t i = 0; i < pad_rows; i++)
uint8_t *src_data_uv = src_data + UVPlaneOffset;
uint8_t *src_data_uv_end = src_data_uv + uv_plane_size;
for (uint32_t i = 0; i < pad_rows / 2; i++)
{
MOS_SecureMemcpy(src_data_uv_end + i * pitch, pitch, src_data_uv_end - pitch, pitch);
}

m_osInterface->pfnUnlockResource(m_osInterface, &(psSurface->OsResource));
}
else
{
MOS_SecureMemcpy(src_data_uv_end + i * uv_pitch, uv_pitch, src_data_uv_end - uv_pitch, uv_pitch);
// we don't copy out the whole tiled buffer to linear and padding on the tiled buffer directly.
lockFlags.TiledAsTiled = 1;

uint8_t *src_data = (uint8_t *)m_osInterface->pfnLockResource(m_osInterface, &(psSurface->OsResource), &lockFlags);
if (!src_data)
return;

uint8_t* padding_data = (uint8_t *)MOS_AllocMemory(pitch * pad_rows);

// Copy last Y row data to linear padding data.
GMM_RES_COPY_BLT gmmResCopyBlt = {0};
gmmResCopyBlt.Gpu.pData = src_data;
gmmResCopyBlt.Gpu.OffsetX = 0;
gmmResCopyBlt.Gpu.OffsetY = (YPlaneOffset + y_plane_size - pitch) / pitch;
gmmResCopyBlt.Sys.pData = padding_data;
gmmResCopyBlt.Sys.RowPitch = pitch;
gmmResCopyBlt.Sys.BufferSize = pitch * pad_rows;
gmmResCopyBlt.Sys.SlicePitch = pitch;
gmmResCopyBlt.Blt.Slices = 1;
gmmResCopyBlt.Blt.Upload = false;
gmmResCopyBlt.Blt.Width = pitch;
gmmResCopyBlt.Blt.Height = 1;
psSurface->OsResource.pGmmResInfo->CpuBlt(&gmmResCopyBlt);
// Fill the remain padding lines with last Y row data.
for (uint32_t i = 1; i < pad_rows; i++)
{
MOS_SecureMemcpy(padding_data + i * pitch, pitch, padding_data, pitch);
}
// Filling the padding for Y.
gmmResCopyBlt.Gpu.pData = src_data;
gmmResCopyBlt.Gpu.OffsetX = 0;
gmmResCopyBlt.Gpu.OffsetY = (YPlaneOffset + y_plane_size) / pitch;
gmmResCopyBlt.Sys.pData = padding_data;
gmmResCopyBlt.Sys.RowPitch = pitch;
gmmResCopyBlt.Sys.BufferSize = pitch * pad_rows;
gmmResCopyBlt.Sys.SlicePitch = pitch;
gmmResCopyBlt.Blt.Slices = 1;
gmmResCopyBlt.Blt.Upload = true;
gmmResCopyBlt.Blt.Width = pitch;
gmmResCopyBlt.Blt.Height = pad_rows;
psSurface->OsResource.pGmmResInfo->CpuBlt(&gmmResCopyBlt);

// Copy last UV row data to linear padding data.
gmmResCopyBlt.Gpu.pData = src_data;
gmmResCopyBlt.Gpu.OffsetX = 0;
gmmResCopyBlt.Gpu.OffsetY = (UVPlaneOffset + uv_plane_size - pitch) / pitch;
gmmResCopyBlt.Sys.pData = padding_data;
gmmResCopyBlt.Sys.RowPitch = pitch;
gmmResCopyBlt.Sys.BufferSize = pitch * pad_rows / 2;
gmmResCopyBlt.Sys.SlicePitch = pitch;
gmmResCopyBlt.Blt.Slices = 1;
gmmResCopyBlt.Blt.Upload = false;
gmmResCopyBlt.Blt.Width = pitch;
gmmResCopyBlt.Blt.Height = 1;
psSurface->OsResource.pGmmResInfo->CpuBlt(&gmmResCopyBlt);
// Fill the remain padding lines with last UV row data.
for (uint32_t i = 1; i < pad_rows / 2; i++)
{
MOS_SecureMemcpy(padding_data + i * pitch, pitch, padding_data, pitch);
}
// Filling the padding for UV.
gmmResCopyBlt.Gpu.pData = src_data;
gmmResCopyBlt.Gpu.OffsetX = 0;
gmmResCopyBlt.Gpu.OffsetY = (UVPlaneOffset + uv_plane_size) / pitch;
gmmResCopyBlt.Sys.pData = padding_data;
gmmResCopyBlt.Sys.RowPitch = pitch;
gmmResCopyBlt.Sys.BufferSize = pitch * pad_rows / 2;
gmmResCopyBlt.Sys.SlicePitch = pitch;
gmmResCopyBlt.Blt.Slices = 1;
gmmResCopyBlt.Blt.Upload = true;
gmmResCopyBlt.Blt.Width = pitch;
gmmResCopyBlt.Blt.Height = pad_rows / 2;
psSurface->OsResource.pGmmResInfo->CpuBlt(&gmmResCopyBlt);

MOS_FreeMemory(padding_data);
padding_data = nullptr;
m_osInterface->pfnUnlockResource(m_osInterface, &(psSurface->OsResource));
}

m_osInterface->pfnUnlockResource(m_osInterface, &(psSurface->OsResource));
}
}

Expand Down
2 changes: 1 addition & 1 deletion media_softlet/linux/Xe_M_plus/ddi/media_sku_wa_mtl.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -284,7 +284,7 @@ static bool InitMtlMediaWaExt(struct GfxDeviceInfo *devInfo,

MEDIA_WR_WA(waTable, WaDisableSetObjectCapture, 1);

MEDIA_WR_WA(waTable, Wa_Vp9UnalignedHeight, 0);
MEDIA_WR_WA(waTable, Wa_Vp9UnalignedHeight, 1);

MEDIA_WR_WA(waTable, Wa_15013355402, 1);

Expand Down

0 comments on commit 1c75882

Please sign in to comment.